The point of this particular discussion is to understand whether specification of RADIUS transport behavior might be helpful in that regard, or whether we've have to go further (such as specifying failover behavior). Reading RFC 2865, one of the reasons that RADIUS was not made to run over TCP in the first place was that it was desired that failover occur in a timeframe prior to connection failure. RFC 3539 handles this issue by defining application-layer timers and heartbeats that allow the AAA application to re-send an accounting packet over another connection before tearing down a suspect one. Via the heartbeat mechanism, the AAA client can determine whether the issue is due to its immediate connection, or something downstream. Failover only occurs if the immediate connection is found to be suspect, so failover occurs on a hop-by-hop basis. RFC 3539 cannot be applied to the RADIUS protocol as it stands because RADIUS does not support a heartbeat. As a result, a RADIUS client can failover even where its immediate proxy is healthy, because of a problem on a downstream RADIUS server. Since RADIUS failover is typically end-to-end, there may be no failover in proxies, even if a server is not responding to requests proxied to it. Even if one were to define appropriate RTT/RTO measurements, and use traffic from the proxy as a demonstration of "liveness", inappropriate failover can still happen in cases where the time between proxy traffic is greater than the failover timer. The end result is that to be able to apply RFC 3539, one would need to add a heartbeat command to RADIUS. This is a fairly major step, since it would require changing both RADIUS client and server implementations. So the question is: "If a major protocol change were to be made to RADIUS to improve reliability, would such a change be deployed on RADIUS clients and servers and would it be acceptable for the SIP accounting specification to depend on such a change?" If the answer is yes, then a RADIUS failover spec might be worth discussing further. If no, then it seems to me that a RADIUS failover spec it might not be worth doing -- it's not considered important enough to get over the deployment hurdles. That raises the question of whether a dependency even on transport behavior improvements is possible, since even this would still require changes on RADIUS clients (though not servers). Comments?